How the X (Twitter) algorithm works, according to its own code

X has published its ranking code twice, and the 2023 version is outdated

Few platforms publish the code behind their main feed. X has. On 31 March 2023, Twitter released the source code of its recommendation algorithm, which picks posts for the For You timeline. confirmed · 2, 3 It came as two GitHub repositories, with an engineering blog post explaining it. confirmed · 1, 3, 4 Twitter said it left out code that could put user safety and privacy at risk or help bad actors, plus the ad recommendation code, training data and model weights. confirmed · 2 The main repository was updated until July 2023, then once more on 3 September 2025. confirmed · 3

Today's system is a separate codebase, published by xAI on GitHub as x-algorithm: "the core code that determines which posts a viewer sees in the For You feed on X". confirmed · 5 Its settings file was last synced with X's configuration system on 18 September 2026. confirmed · 5 X calls these settings its "primary production values". confirmed · 5

Mind the date gap. Many explainers still rely on the March 2023 release, which describes the Twitter-era system: about 1,500 candidate posts per request, and a timeline that was on average half posts from accounts you follow. confirmed · 1 The 2026 code uses a different programming language, a different model and a different list of predicted actions. confirmed · 3, 5 Even X's own Help Center page on For You is undated and still links to the 2023 repository. confirmed · 6

Today's feed predicts two dozen reactions per post, then weighs them

1Understanding posts and accounts. A classifier called Grox sorts new posts into categories such as spam, adult content and violent media, and turns their text and images into numeric summaries. confirmed · 5 Separate models check images and video. confirmed · 5 Other jobs rate accounts: one labels an account by how people respond to its posts (blocks, reports and spam reports, relative to likes), one looks for inauthentic behavior, and one computes a reputation score from follows and engagement. confirmed · 5 These labels decide whether a post can be shown at all, not where it ranks. confirmed · 5

2Finding candidates (possible posts). Each time the app asks for posts, recent posts from accounts you follow come from a live store called Thunder. confirmed · 5 Posts from accounts you don't follow come from Phoenix retrieval, which matches your recent activity against posts described by their content and author, and from SimClusters, which groups accounts and posts by who engages with what. confirmed · 5 To that retrieval model, you're mostly what you've interacted with: it learns no separate profile for you and reads only your history, plus a few details such as country and language. confirmed · 5

3First filters. Before scoring, the code removes duplicates, posts older than 48 hours, your own posts, reposts and replies from accounts you don't follow, posts from accounts you block or mute, posts with your muted keywords, posts you've already seen, and subscriber-only posts you can't open. confirmed · 5

4Ranking by predicted actions. A transformer model called Phoenix, the same family of AI that powers chatbots, reads your recent engagement history. confirmed · 5 It predicts how likely you are to like, reply, repost, quote, share, click, watch, linger, follow the author, or tap "not interested", mute, block or report. confirmed · 5 It also sees context such as the post's age and your local time. confirmed · 5 It scores each post on its own, so other posts in the batch can't change its score. confirmed · 5 Then the predictions are multiplied by weights and added up. confirmed · 5 The 18 September 2026 settings include these weights:

• Reply 5, raised to 20 on original posts from someone you follow who follows you back (a boost X tested from 10 July 2026 and set at this level on 24 July). confirmed · 5
• Quote 5, share 2, share by DM 5, share by copying the link 20, follow the author 4. confirmed · 5
• Repost 1, like 0.5, click into the post 0.4, open a link 0.2, plus small weights for lingering on a post and opening photos and videos. confirmed · 5
• Profile clicks and "video quality views" are predicted but currently weighted zero. confirmed · 5
• Not interested −43.2, mute −58.8, block −31.2, report −234. confirmed · 5

X left a warning in the code: the weights multiply one viewer's predicted probabilities, not counts, so reading them as "one report cancels 468 likes" is wrong. confirmed · 5 Reports weigh heavily because they're more than 1,000 times rarer than likes. confirmed · 5

Three adjustments follow. Posts from accounts you don't follow, and replies and reposts from accounts you do follow, are multiplied by 0.75. confirmed · 5 Twitter's 2023 code did the same for accounts you don't follow. confirmed · 3 Extra posts from one author in the same refresh are discounted more and more, but never below a quarter of their score. confirmed · 5 By my arithmetic with these settings, the author's second post keeps about 62% of its score and the third about 44%. inferred · 5 A new-author boost can lift one post per refresh to around 16th place, if it's an original post under 48 hours old, from an account with 1,000 or fewer followers, with fewer than 1,000 views on the Home timeline. confirmed · 5 A last step spreads out posts that look too alike, giving up a little score for variety. confirmed · 5

For comparison, the 2023 ranking model had about 48 million parameters and predicted ten kinds of engagement, including likes, retweets, replies, replies the author engaged with, and reports. confirmed · 1, 3

5Final checks and mixing. After ranking, a visibility service decides for each post and viewer: allow, show behind a warning screen, or drop. confirmed · 5 Some drop rules apply only when a post would be recommended to someone who doesn't follow the author, such as rules for "do not amplify" labels, a broad spam check, malicious links and abuse. confirmed · 5 Followers can still see the same post. confirmed · 5 Duplicate branches of a conversation are collapsed, and ads, Who to Follow and prompts are mixed in. confirmed · 5

6Learning from what happens next. Your recent actions are the main input the model reads on your next request. confirmed · 5 So what you engage with today shapes what For You shows you tomorrow. inferred · 5 The repository has the code that trains the production model, but not the trained model, and doesn't say how often the live model is retrained. confirmed · 5 X's Help Center only says it's "continuously trained on your interactions". confirmed · 6

What this means for your posts: earn replies and shares, not mutes

My deductions from the code, not X's advice. They assume the September 2026 settings, which X can change at any time. confirmed · 5

• Write to be answered. A predicted reply or quote weighs ten times a like, and a reply forty times on original posts between mutual follows. confirmed · 5 So a post with something specific to answer, like a question or a clear claim, should outrank one that only collects likes. inferred
• Make something worth sending. A share by copied link has the heaviest base weight, forty times a like, and a DM share ten times. confirmed · 5 Posts people pass on to a friend or a group chat, like a clear explainer or a useful list, likely gain more than posts they merely agree with. inferred
• Don't earn the mute. Predicted "not interested" taps, mutes, blocks and reports carry large negative weights, and account labels are built partly from blocks and reports relative to likes. confirmed · 5 Rage bait that draws replies from a few and mutes from many likely loses more than it gains, and risks labels that stop recommendations to non-followers. inferred
• Treat every post as a two-day window. For You drops posts older than 48 hours, and the model sees the viewer's local time. confirmed · 5 So post when your audience is active, and expect later reach to come from outside For You, such as search or your profile. inferred
• Space your posts out. When several of your posts are candidates in one follower's refresh, each after the first is discounted more than the last. confirmed · 5 Spread them out and each can reach that follower in a fresh refresh, where posts they've already seen are filtered out. inferred
• Small account? Post originals. The new-author boost only considers original posts from accounts with 1,000 or fewer followers, and For You filters out replies from accounts the viewer doesn't follow. confirmed · 5 Replies under big accounts can get you seen in those threads, but your own originals are what For You can test with new people. inferred

What nobody outside X can know, and four beliefs I checked

The code shows the rules, not the judgments. The trained model isn't published, and its predictions depend on each viewer's history. confirmed · 5 So no one outside X can compute the score your next post will get. inferred · 5 I watch videos all day, and I can't either.

X says it keeps the Grox classifier's prompts and some labeling rules out of the code so the system is harder to game. confirmed · 5 Settings can differ for people inside experiments, and X only aims to publish experiments that reach 10% or more of traffic. confirmed · 5 The repository focuses on For You and doesn't cover how replies are ordered under a post, or how Search and Explore rank. confirmed · 5

• "One report cancels out hundreds of likes." Contradicted: the weights apply to predicted probabilities, not counts. confirmed · 5 X's code comments add that reports from bad actors mostly affect recommendations for users similar to them, and that only actions on posts served in the Home timeline count for ranking. confirmed · 5 Blocks and reports do feed account-level labels, though. confirmed · 5 How much a coordinated campaign could move those labels isn't published.
• "Premium boosts your posts." It did in the 2023 code, not in today's For You scoring. confirmed · 3, 5 The March 2023 code multiplied "Blue Verified" authors' scores by default factors of 4 for their followers and 2 for everyone else. confirmed · 3 Those multipliers are gone from the 2025 update. confirmed · 3 X's Premium page does promise "reply prioritization" in conversations. confirmed · 7 And verified accounts are part of the starting point for the reputation score that labeling rules read. confirmed · 5
• "Links kill your reach." I found no published paper, code or official page that confirms a penalty on ordinary links. In the 2026 settings, opening a link has a small positive weight, and the one link rule I found drops recommended posts with malicious links. confirmed · 5 The 2023 to 2025 code also removed posts from accounts you don't follow if they linked to sites on a "competitor" list, but the list itself wasn't in the code. confirmed · 3
• "X shadowbans accounts." Confirmed in a narrower form: labels such as "do not amplify", or the broad spam check, can stop a post being recommended to non-followers while followers still see it. confirmed · 5 X's Under the Hood tool, still a pilot, shows statistics on these labels for your account and posts. confirmed · 5